Research Driving the Future of AI in Manufacturing

Universities play an important role in advancing applications for AI into the manufacturing sector.  At the forefront of many of these advances is Carnegie Mellon University (CMU), particularly well known for its work in AI, machine learning (ML), data analytics, robotics, and metals additive manufacturing. 

CMU’s Manufacturing Futures Institute plays a key role in catalyzing this research through its mission to advance technologies and workforce development toward more agile, intelligent, efficient, resilient and sustainable manufacturing.  In this talk, I provide an overview of a sampling of recent work undertaken by my colleagues on AI and ML related to the future of manufacturing.

ML is often associated with creating surrogate models that run fast, for example for real time analytic or control applications. Fast models also make feasible design optimization where reaching a desired solution may take many thousands of computational iterations. At CMU, additive manufacturing is one area of emphasis for exploring such ML use. Some exemplary projects are metal melt pool shape, depth and temperature prediction from surface thermal images; predicting part porosity based on thermal images; residual deformation learning and mitigation for metal component printability enhancement; and inverse design and manufacture of self-assembling fiber-reinforced composites.

Robotics with judicious use of AI can provide solutions to manufacturing steps that are tedious, difficult and dangerous for workers to do. An example is robotic grinding of weld seams in tight spaces, where a combination of self-calibration, efficient motion planning, and hybrid force-position control enabled a robot to complete the task successfully. A more futuristic robot capability is the ability to efficiently manipulate unseen objects and transfer relevant skills to new use cases.

A relatively new aspect of digital transformation are digital twins, which will eventually impact all manufacturing sectors.  The synergies between digital twins and AI are just beginning to be recognized. Incorporating the physical twin’s data to update the digital twin provides the information and context of the physical world to AI applications. Some clues come from roboticists, who use virtual models regularly in training, and from civil engineers, who use digital twins to interact with infrastructure systems.

I will conclude by touching upon large language models and diffusion models, which are making an impact across sectors, including in manufacturing. Applications range from help with coding, to generative models for manufacturing, to more fundamental use in finding physics-based models from data.